Hybrid offline-online machine learning framework for real-time UAV battery voltage prediction

Unmanned Aerial Vehicles (UAVs) are increasingly adopted in logistics and mission-critical applications, where reliable battery state-of-energy information is essential for safe planning and execution. However, robust voltage prediction for real-time operations remains underexplored, and existing approaches have not explored the possibility of combining offline prediction with discrepancy-triggered in-flight updates to handle the uncertainty raising from the lack of knowledge about future power profiles. In this sense, this paper proposes a dual-phase voltage prediction framework that (i) provides a first estimate of the voltage trend along a planned route using an offline multi-criteria segmentation of historical flights and a route-matching procedure, and (ii) refines predictions in-flight through a discrepancy-triggered update based on the Continuous Ranked Probability Score (CRPS), rerunning the model only when significant deviations are detected to limit computational burden and to adapt to unseen mission profiles. The method combines a hybrid physics-based/data-driven formulation with a Gated Recurrent Unit (GRU) trained on up to 200 flights to model battery behavior and to forecast the full voltage trajectory. Validation on a publicly available dataset shows that the proposed offline-online strategy substantially improves prediction accuracy, reducing MAE from 0.181 V to 0.087 V (51 % improvement). A sensitivity analysis on the CRPS threshold parameters highlights the trade-off between error reduction and computational cost. The overall work incorporates innovative methodologies with a focus on practical applicability, thus facilitating the advancement of ML strategies in these domains.